CN108805152A - A kind of scene classification method and device - Google Patents
A kind of scene classification method and device Download PDFInfo
- Publication number
- CN108805152A CN108805152A CN201710313796.XA CN201710313796A CN108805152A CN 108805152 A CN108805152 A CN 108805152A CN 201710313796 A CN201710313796 A CN 201710313796A CN 108805152 A CN108805152 A CN 108805152A
- Authority
- CN
- China
- Prior art keywords
- scene
- multiple dimensioned
- scale
- neural networks
- convolutional neural
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
Landscapes
- Engineering & Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Theoretical Computer Science (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Artificial Intelligence (AREA)
- Evolutionary Biology (AREA)
- Evolutionary Computation (AREA)
- Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Image Analysis (AREA)
Abstract
A kind of scene classification method of present invention offer and device, including：S1, multiple dimensioned convolutional neural networks are based on, extract scene convolution feature of the scene picture of input on each scale；S2, Fusion Features are carried out to the scene convolution feature on each scale, obtains the multiple dimensioned scene characteristic of the scene picture；S3, it is based on the multiple dimensioned scene characteristic, scene classification is completed in the multiple dimensioned convolutional neural networks.A kind of scene classification method proposed by the present invention and device have fully excavated the contact of scene characteristic between different scale, have extracted the multiple dimensioned scene characteristic with identification, improve the precision of scene classification by building multiple dimensioned convolutional neural networks.
Description
Technical field
The present invention relates to information technology fields, more particularly, to a kind of scene classification method and device.
Background technology
With the rapid development of the multimedia technologies such as Digital photographic and stored digital, the digital image data that people obtain is got over
Come more, contains magnanimity information in these image datas, only relying on manpower at all can not locate these magnanimity informations in real time
Reason.Therefore, it is desirable to which the Vision information processing function by simulating human visual system, assigns the energy of machine automatic identification image
Power, to help or the mankind is assisted to complete many vital tasks.Scene classification is inferred to this by the content that picture is included
The correct classification of scene is a basic and very challenging computation vision task, in picture retrieval, target detection
And the fields such as target following play an important role.
Currently, whether existing scene recognition method is related to convolutional neural networks according to model used, it is broadly divided into two classes：
One type is the shallow Model based on manual feature, and such methods are dedicated to designing the scene characteristic operator of robust, or set
Count the model of robust；Another type is the depth model based on convolutional neural networks, and the core concept of such methods is will to roll up
Product neural network extracts the scene characteristic for including high-rise semantics information, classifies as scene characteristic extractor.
But based on the depth model of convolutional neural networks by feature learning and classifier training point in processing procedure
From weakening the performance of entire model, and the model, by the feature under different scale, directly fusion obtains scene characteristic, not
Give full play to the performance of scene characteristic in single scale so that nicety of grading is not high.
Invention content
The present invention provides a kind of a kind of scene classification side for overcoming the above problem or solving the above problems at least partly
Method and device.
According to the first aspect of the invention, a kind of scene classification method is provided, including：
S1, multiple dimensioned convolutional neural networks are based on, it is special extracts scene convolution of the scene picture of input on each scale
Sign；
S2, Fusion Features are carried out to the scene convolution feature on each scale, obtains more rulers of the scene picture
Spend scene characteristic；
S3, it is based on the multiple dimensioned scene characteristic, scene classification is completed in the multiple dimensioned convolutional neural networks.
Wherein, the multiple dimensioned convolutional neural networks include convolutional layer, full articulamentum, activation primitive layer, SoftMax layers with
And on the convolutional layer layer building multiple dimensioned layer.
Wherein, step S2 includes：
S21, pre-confluent is carried out to the scene convolution feature in each scale, it is special obtains scene fusion in each scale
Sign；
S22, Fusion Features are carried out to scene fusion feature in each scale between each scale, obtains the field
The multiple dimensioned scene characteristic of scape picture.
Wherein, fusion process is all made of ReLU activation primitives and is merged to feature in step S21 and step S22.
Wherein, step S1 includes：
The scene picture of input is divided into the scenario block of multiple scales in multiple dimensioned layer；
The scene convolution feature of the scenario block of each scale is extracted in convolutional layer.
Wherein, step S3 includes：
Object function is built in the multiple dimensioned convolutional neural networks；
Based on the multiple dimensioned scene characteristic and the object function, classify in the SoftMax layer buildings scene characteristic
Device completes scene classification.
Wherein, the object function is：
Wherein, the scene number of pictures that { x, y } is the scene picture x of input and its label y, M are input, C is scene
Class number, RmsFor multiple dimensioned scene characteristic, the weight and biasing of W and b for multiple dimensioned convolutional neural networks grader.
Wherein, it is described to the multiple dimensioned convolutional neural networks be trained including：
The multiple dimensioned convolutional neural networks are trained using the stochastic gradient descent method with momentum, until the mesh
Scalar functions are restrained.
According to the second aspect of the invention, a kind of scene classification device is provided, including：
Extraction module extracts the scene picture of input on each scale for being based on multiple dimensioned convolutional neural networks
Scene convolution feature；
Fusion Module obtains the scene for carrying out Fusion Features to the scene convolution feature on each scale
The multiple dimensioned scene characteristic of picture；
Sort module completes field for being based on the multiple dimensioned scene characteristic in the multiple dimensioned convolutional neural networks
Scape is classified.
According to the third aspect of the invention we, a kind of computer program product, including program code, said program code are provided
For executing image search method described above.
According to the fourth aspect of the invention, a kind of non-transient computer readable storage medium is provided, for storing such as preceding institute
The computer program stated.
A kind of scene classification method proposed by the present invention and device are fully sent out by building multiple dimensioned convolutional neural networks
The contact of scene characteristic between different scale has been dug, the multiple dimensioned scene characteristic with identification has been extracted, improves scene classification
Precision.
Description of the drawings
Fig. 1 is a kind of scene classification method flow chart provided in an embodiment of the present invention；
Fig. 2 is another scene classification method flow chart provided in an embodiment of the present invention；
Fig. 3 is a kind of scene classification structure drawing of device provided in an embodiment of the present invention.
Specific implementation mode
With reference to the accompanying drawings and examples, the specific implementation mode of the present invention is described in further detail.Implement below
Example is not limited to the scope of the present invention for illustrating the present invention.
Fig. 1 is a kind of scene classification method flow chart provided in an embodiment of the present invention, as shown in Figure 1, the method includes：
S1, multiple dimensioned convolutional neural networks are based on, it is special extracts scene convolution of the scene picture of input on each scale
Sign；
S2, Fusion Features are carried out to the scene convolution feature on each scale, obtains more rulers of the scene picture
Spend scene characteristic；
S3, it is based on the multiple dimensioned scene characteristic, scene classification is completed in the multiple dimensioned convolutional neural networks.
In S1, the multiple dimensioned convolutional neural networks are the multiple dimensioned convolutional neural networks that training is completed, specifically, in advance
Multiple dimensioned convolutional neural networks are built, the input multiple dimensioned convolutional neural networks of training sample set pair are trained, after training
To the neural network structure with scene classification function.
It is understood that in the classification for the training sample set that the classification of the scene picture of input need to be included in input.
In S2, Fusion Features are carried out to the scene convolution feature on each scale and use secondary convergence strategy, in ruler
It in degree on the basis of pre-confluent, then carries out second between scale and merges, obtain the multiple dimensioned scene characteristic of scene picture, this is more
Scale scene characteristic identification is strong, and has very strong robustness to geometry rotation, and efficiently solving part scene picture has
The problem of blocking.
In S3, scene classification completion scene classification is completed directly in multiple dimensioned convolutional neural networks, is no longer needed to by outer
In grader, the working performance of multiple dimensioned convolutional neural networks is given full play to, scene point is completed while carrying out feature extraction
Class.
Scene classification method provided in an embodiment of the present invention is fully excavated by building multiple dimensioned convolutional neural networks
The contact of scene characteristic between different scale extracts the multiple dimensioned scene characteristic with identification, improves the essence of scene classification
Degree.
On the basis of the above embodiments, multiple dimensioned convolutional neural networks provided in an embodiment of the present invention include：
Convolutional layer, full articulamentum, activation primitive layer and SoftMax layers, and the more rulers of layer building on the convolutional layer
Spend layer.
In general, the basic structure of traditional convolutional neural networks includes two layers, one is characterized extract layer, i.e. convolutional layer,
Each input of neuron is connected with the local acceptance region of preceding layer, and extracts the feature of the part.Once the local feature quilt
After extraction, its position relationship between other feature is also decided therewith；The second is Feature Mapping layer, includes：Full connection
Layer, activation primitive layer and SoftMax layers.Each computation layer of network is made of multiple Feature Mappings, and each Feature Mapping is
One plane, the weights of all neurons are equal in plane.
Wherein, the multiple dimensioned convolutional neural networks are embedded in one on the convolutional layer upper layer of traditional convolutional neural networks
Multiple dimensioned layer, after scene picture is input into the multiple dimensioned layer, by according to the default segmentation scale of scale layer, to scene graph
Piece is split.
Described in Fig. 1 on the basis of embodiment, Fig. 2 is another scene classification method flow provided in an embodiment of the present invention
Figure, as shown in Fig. 2, step S2 includes：
S21, pre-confluent is carried out to the scene convolution feature in each scale, it is special obtains scene fusion in each scale
Sign；
S22, Fusion Features are carried out to scene fusion feature in each scale between each scale, obtains the field
The multiple dimensioned scene characteristic of scape picture.
In S21, the convolution feature of each scenario block is not single features, the volume of each scenario block on each scale
Product feature is possible to more than one, in order to fully excavate the performance of feature in scale, thus it is pre- to convolution feature in each scale
Fusion, calculating process are represented by：
Rl=σ (Wfl[σ(Wclrl,1+bcl),…σ(Wclrl,16+bcl)]+bfl)
Rm=σ (Wfm[σ(Wcmrm,1+bcm),…σ(Wcmrl,4+bcm)]+bfm)
Rg=σ (Wfgrg+bfg)
Wherein, σ () is Rectified Linear Units (ReLU) activation primitive, rl,i、rm,i、rgFor different scale
Under each scenario block convolution feature, Rl、Rm、RgFor scene fusion feature in each scale, [A, B] indicates A connecting shape with B
The matrix of Cheng Xin.
It in S22, obtains in each scale after scene fusion feature, Fusion Features is carried out to scene fusion feature, are had
There are the multiple dimensioned scene characteristic of identification, calculating process to be represented by：
Rms=σ (Wms[Rl,Rm,Rg]+bms)
Wherein, RmsFor multiple dimensioned scene characteristic, WmsAnd bmsIndicate weight and the biasing of multiple dimensioned scene characteristic.
The embodiment of the present invention carries out pre-confluent by the convolution feature to the scenario block in scale, fully excavates on each scale
Association between feature, the Fusion Features between scale provide better basis.
On the basis of the above embodiments, fusion process is all made of ReLU activation primitives to spy in step S21 and step S22
Sign is merged.
The ReLu activation primitives can greatly accelerate convergence rate so that feature is melted during Fusion Features
The problem of closing faster, and alleviating gradient disperse so that gradient will not be saturated quickly, and fusion is newer more efficient.
Described in Fig. 1 on the basis of embodiment, step S1 includes：
The scene picture of input is divided into the scenario block of multiple scales in multiple dimensioned layer；
The scene convolution feature of the scenario block of each scale is extracted in convolutional layer.
Wherein, a multiple dimensioned layer is established in convolutional neural networks, when scene picture inputs, multiple dimensioned layer is preset
The scene picture automatic cutting of input is to correspond to the scale of scale layer, such as formula by scale：
E (x)={ l1,…l16,m1,…m4,g}
Wherein, x is the scene picture of input, and e () is multiple dimensioned operation, liFor the scenario block on 86 × 86 scales, miFor
Scenario block on 140 × 140 scales, g are the scenario block on 224 × 224 scales.
It should be noted that the embodiment of the present invention is not limited specific scale, the above-mentioned scale provided is only to refer to
Scale.
Wherein, convolution module is established in multiple dimensioned convolutional neural networks, the field of the scenario block for extracting each scale
Convolution mould of all convolutional layers of GoogLeNet as the multiple dimensioned convolutional neural networks of the present invention is employed herein in scape block feature
Block extracts scene block feature respectively on each scale：
rl,i=GoogLeNet (li)
rm,i=GoogLeNet (mi)
rg=GoogLeNet (g)
Wherein, rl,i、rm,i、rgFor the convolution feature of each scenario block under different scale.
The embodiment of the present invention carries out pre-confluent by the convolution feature to the scenario block in scale, fully excavates on each scale
Association between feature, the Fusion Features between scale provide better basis.
Described in Fig. 1 on the basis of embodiment, step S3 includes：
Object function is built in the multiple dimensioned convolutional neural networks；
Based on the multiple dimensioned scene characteristic and the object function, classify in the SoftMax layer buildings scene characteristic
Device completes scene classification.
Multiple dimensioned convolutional neural networks grader builds grader on Feature Mapping layer and classifies to picture, generally
, we use grader of the softmax graders as multiple dimensioned convolutional neural networks.
On the basis of above-described embodiment, the object function is：
Wherein, the scene number of pictures that { x, y } is the scene picture x of input and its label y, M are input, C is scene
Class number, RmsFor multiple dimensioned scene characteristic, the weight and biasing of W and b for multiple dimensioned convolutional neural networks grader.
According to the scene type number for the scene picture that the object function is inputted, the classification of the same category number is built
Device, and the weight of the multiple dimensioned convolutional neural networks grader controlled according to object function and biasing adjust in the training process
Multiple dimensioned scene characteristic, until when object function convergence, the multiple dimensioned scene characteristic is optimal.
The embodiment of the present invention is appointed by multiple dimensioned convolutional neural networks direct construction grader, realizing in scene classification
Multitask strategy in business enhances the performance of entire multiple dimensioned convolutional neural networks structure.
On the basis of the above embodiments, it is described to the multiple dimensioned convolutional neural networks be trained including：
The multiple dimensioned convolutional neural networks are trained using the stochastic gradient descent method with momentum, until the mesh
Scalar functions are restrained.
The stochastic gradient descent method with momentum is to randomly select some training datasets to replace entire sample training
Collection, makees gradient decline on the training dataset randomly selected, and until object function is restrained, the advantage of this method is
Iteration time is saved, improves training effectiveness so that object function being capable of more rapid convergence.
The embodiment of the present invention by using the stochastic gradient descent method with momentum to the multiple dimensioned convolutional neural networks into
Row training so that training speed is accelerated, training effectiveness higher.
Specifically, being extracted first to the multiple dimensioned scene characteristic of the scene picture of input, further according to more rulers of extraction
Degree scene characteristic classifies the scene picture of input, and in order to further verify using multiple dimensioned scene characteristic, the present invention is real
Example is applied to emulate above-mentioned scene classification method.
1. simulated conditions
It is Intel (R) Core i7-5930K3.50GHZ, GeForce GTX that the embodiment of the present invention, which is in central processing unit,
On Titan X GPU, memory 64G, linux operating system, with the emulation of MATLAB softwares progress.
What emulation experiment data utilized is by the Massachusetts Institute of Technology (Massachu-setts Institute of
Technology, MIT) the J.Xiao et al. of 67 databases of MIT indoor and MIT that provides of A.Quattoni et al.
397 databases of SUN of offer.
2. emulation content
The picture random division 80% in database that emulation is used is used as training sample set, and remaining 20% as survey
Sample set is tried, training sample set is input in the multiple dimensioned convolutional neural networks built, to multiple dimensioned convolutional neural networks
It is trained, input test sample set after training, and classifies to test sample collection.By obtained classification results and test specimens
The legitimate reading of this collection compares, the correct number r of statistical classification, and nicety of grading is then：
Acc=r/R*100%
Wherein, R is the number of samples of test sample collection.
Multiple dimensioned disordering neural network (MOP-CNN), simple files check addition are used respectively with same database
(SFV), oriented acyclic neural network (DAG-CNN), Digital Signal Processing method (DSP) are tested, and count this 5 kinds of methods
To the nicety of grading of the database.
Using the niceties of grading of 67 databases of MIT indoor, the results are shown in Table 1：
Table 1：Scene nicety of grading on 67 databases of MIT indoor
| Sorting technique | Nicety of grading |
| MOP-CNN | 68.88% |
| SFV | 72.86% |
| DAG-CNN | 77.50% |
| DSP | 78.28% |
| The present invention | 80.90% |
Using the niceties of grading of SUN397 databases, the results are shown in Table 2：
Table 2：Scene nicety of grading on SUN397 databases
| Sorting technique | Nicety of grading |
| MOP-CNN | 51.98% |
| SFV | 54.40% |
| DAG-CNN | 56.20% |
| DSP | 59.78% |
| The present invention | 62.24% |
The scene classification side provided in an embodiment of the present invention it can be seen from the emulation data of the nicety of grading of Tables 1 and 2
Method has in nicety of grading and is obviously improved, this is because the present invention, on the basis of original convolutional neural networks, structure is more
Scale layer, and the scene characteristic between each scale layer merges, and obtains the multiple dimensioned scene characteristic with identification, makes
Nicety of grading is obtained to be obviously improved.
Fig. 3 is a kind of scene classification device provided in an embodiment of the present invention, including：
Extraction module 1 extracts the scene picture of input on each scale for being based on multiple dimensioned convolutional neural networks
Scene convolution feature；
Fusion Module 2 obtains the scene for carrying out Fusion Features to the scene convolution feature on each scale
The multiple dimensioned scene characteristic of picture；
Sort module 3 completes field for being based on the multiple dimensioned scene characteristic in the multiple dimensioned convolutional neural networks
Scape is classified.The embodiment of the present invention additionally provides a kind of storage device, and wherein that is stored with a plurality of instruction, described instruction be suitable for by
Processor is loaded and is executed：
Wherein extraction module 1 is based on trained multiple dimensioned convolutional neural networks, extracts the scene picture of input each
Scene convolution feature on scale.
Wherein, Fusion Module 2 carries out Fusion Features to the scene convolution feature on each scale and uses secondary fusion
Strategy in scale on the basis of pre-confluent, then carries out second between scale and merges, obtains the multiple dimensioned scene of scene picture
Feature, the multiple dimensioned scene characteristic identification is strong, and has very strong robustness to geometry rotation, efficiently solves part
The problem of scape picture blocks.
Wherein, sort module 3 completes scene classification directly in multiple dimensioned convolutional neural networks and completes scene classification, is not necessarily to
External grader is relied on again, gives full play to the working performance of multiple dimensioned convolutional neural networks, it is complete while carrying out feature extraction
At scene classification.
Scene classification method provided in an embodiment of the present invention is filled by building extraction module, Fusion Module and sort module
The contact of scene characteristic between different scale has been dug in distribution, is extracted the multiple dimensioned scene characteristic with identification, is improved scene
The precision of classification.
The present embodiment provides a kind of scene classification devices, including：At least one processor；And with the processor communication
At least one processor of connection, wherein：
The memory is stored with the program instruction that can be executed by the processor, and the processor calls described program to refer to
It enables to execute the method that above-mentioned each method embodiment is provided, such as including：Based on multiple dimensioned convolutional neural networks, extraction input
Scene convolution feature of the scene picture on each scale；Feature is carried out to the scene convolution feature on each scale to melt
It closes, obtains the multiple dimensioned scene characteristic of the scene picture；Based on the multiple dimensioned scene characteristic, in the multiple dimensioned convolution god
Scene classification is completed in network.
The present embodiment discloses a kind of computer program product, and the computer program product includes being stored in non-transient calculating
Computer program on machine readable storage medium storing program for executing, the computer program include program instruction, when described program instruction is calculated
When machine executes, computer is able to carry out the method that above-mentioned each method embodiment is provided, such as including：Based on multiple dimensioned convolution god
Through network, scene convolution feature of the scene picture of input on each scale is extracted；To the scene volume on each scale
Product feature carries out Fusion Features, obtains the multiple dimensioned scene characteristic of the scene picture；Based on the multiple dimensioned scene characteristic,
Scene classification is completed in the multiple dimensioned convolutional neural networks.
The present embodiment provides a kind of non-transient computer readable storage medium, the non-transient computer readable storage medium
Computer instruction is stored, the computer instruction makes the computer execute the method that above-mentioned each method embodiment is provided, example
Such as include：Based on multiple dimensioned convolutional neural networks, scene convolution feature of the scene picture of input on each scale is extracted；It is right
Scene convolution feature on each scale carries out Fusion Features, obtains the multiple dimensioned scene characteristic of the scene picture；Base
In the multiple dimensioned scene characteristic, scene classification is completed in the multiple dimensioned convolutional neural networks.
One of ordinary skill in the art will appreciate that：Realize that all or part of step of above method embodiment can pass through
The relevant hardware of program instruction is completed, and program above-mentioned can be stored in a computer read/write memory medium, the program
When being executed, step including the steps of the foregoing method embodiments is executed；And storage medium above-mentioned includes：ROM, RAM, magnetic disc or light
The various media that can store program code such as disk.
Through the above description of the embodiments, those skilled in the art can be understood that each embodiment can
It is realized by the mode of software plus required general hardware platform, naturally it is also possible to pass through hardware.Based on this understanding, on
Stating technical solution, substantially the part that contributes to existing technology can be expressed in the form of software products in other words, should
Computer software product can store in a computer-readable storage medium, such as ROM/RAM, magnetic disc, CD, including several fingers
It enables and using so that a computer equipment (can be personal computer, server or the network equipment etc.) executes each implementation
Method described in certain parts of example or embodiment.
Finally, the present processes are only preferable embodiment, are not intended to limit the scope of the present invention.It is all
Within the spirit and principles in the present invention, any modification, equivalent replacement, improvement and so on should be included in the protection of the present invention
Within the scope of.
Claims (11)
1. a kind of scene classification method, which is characterized in that including：
S1, multiple dimensioned convolutional neural networks are based on, extract scene convolution feature of the scene picture of input on each scale；
S2, Fusion Features are carried out to the scene convolution feature on each scale, obtains the multiple dimensioned field of the scene picture
Scape feature；
S3, it is based on the multiple dimensioned scene characteristic, scene classification is completed in the multiple dimensioned convolutional neural networks.
2. according to the method described in claim 1, the multiple dimensioned convolutional neural networks include convolutional layer, full articulamentum, activation
The multiple dimensioned layer of layer building on function layer, SoftMax layers and the convolutional layer.
3. according to the method described in claim 1, it is characterized in that, step S2 includes：
S21, pre-confluent is carried out to the scene convolution feature in each scale, obtains scene fusion feature in each scale；
S22, Fusion Features are carried out to scene fusion feature in each scale between each scale, obtains the scene graph
The multiple dimensioned scene characteristic of piece.
4. according to the method described in claim 3, it is characterized in that, fusion process is all made of ReLU in step S21 and step S22
Activation primitive merges feature.
5. according to the method described in claim 2, it is characterized in that, step S1 includes：
The scene picture of input is divided into the scenario block of multiple scales in multiple dimensioned layer；
The scene convolution feature of the scenario block of each scale is extracted in convolutional layer.
6. according to the method described in claim 2, it is characterized in that, step S3 includes：
Object function is built in the multiple dimensioned convolutional neural networks；
It is complete in the SoftMax layer buildings scene characteristic grader based on the multiple dimensioned scene characteristic and the object function
At scene classification.
7. according to the method described in claim 6, it is characterized in that, the object function is：
Wherein, the scene number of pictures that { x, y } is the scene picture x of input and its label y, M are input, C is scene type
Number, RmsFor multiple dimensioned scene characteristic, the weight and biasing of W and b for multiple dimensioned convolutional neural networks grader.
8. the method according to the description of claim 7 is characterized in that the method further includes：
The multiple dimensioned convolutional neural networks are trained using the stochastic gradient descent method with momentum, until the target letter
Number convergence.
9. a kind of scene classification device, which is characterized in that including：
Extraction module extracts scene of the scene picture of input on each scale for being based on multiple dimensioned convolutional neural networks
Convolution feature；
Fusion Module obtains the scene picture for carrying out Fusion Features to the scene convolution feature on each scale
Multiple dimensioned scene characteristic；
Sort module completes scene point for being based on the multiple dimensioned scene characteristic in the multiple dimensioned convolutional neural networks
Class.
10. a kind of computer program product, which is characterized in that the computer program product includes being stored in non-transient computer
Computer program on readable storage medium storing program for executing, the computer program include program instruction, when described program is instructed by computer
When execution, the computer is made to execute method as described in any of the claims 1 to 8.
11. a kind of non-transient computer readable storage medium, which is characterized in that the non-transient computer readable storage medium is deposited
Computer instruction is stored up, the computer instruction makes the computer execute method as described in any of the claims 1 to 8.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710313796.XA CN108805152A (en) | 2017-05-05 | 2017-05-05 | A kind of scene classification method and device |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710313796.XA CN108805152A (en) | 2017-05-05 | 2017-05-05 | A kind of scene classification method and device |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN108805152A true CN108805152A (en) | 2018-11-13 |
Family
ID=64054840
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201710313796.XA Pending CN108805152A (en) | 2017-05-05 | 2017-05-05 | A kind of scene classification method and device |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN108805152A (en) |
Cited By (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109670516A (en) * | 2018-12-19 | 2019-04-23 | 广东工业大学 | A kind of image characteristic extracting method, device, equipment and readable storage medium storing program for executing |
| CN109784159A (en) * | 2018-12-11 | 2019-05-21 | 北京航空航天大学 | The processing method of scene image, apparatus and system |
| CN113033527A (en) * | 2021-05-27 | 2021-06-25 | 北京三快在线科技有限公司 | Scene recognition method and device, storage medium and unmanned equipment |
Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN105787510A (en) * | 2016-02-26 | 2016-07-20 | 华东理工大学 | System and method for realizing subway scene classification based on deep learning |
| CN105956532A (en) * | 2016-04-25 | 2016-09-21 | 大连理工大学 | Traffic scene classification method based on multi-scale convolution neural network |
-
2017
- 2017-05-05 CN CN201710313796.XA patent/CN108805152A/en active Pending
Patent Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN105787510A (en) * | 2016-02-26 | 2016-07-20 | 华东理工大学 | System and method for realizing subway scene classification based on deep learning |
| CN105956532A (en) * | 2016-04-25 | 2016-09-21 | 大连理工大学 | Traffic scene classification method based on multi-scale convolution neural network |
Non-Patent Citations (1)
| Title |
|---|
| QINGSHAN LIU 等: "Adaptive Deep Pyramid Matching for Remote Sensing Scene Classification", 《IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING》 * |
Cited By (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109784159A (en) * | 2018-12-11 | 2019-05-21 | 北京航空航天大学 | The processing method of scene image, apparatus and system |
| CN109670516A (en) * | 2018-12-19 | 2019-04-23 | 广东工业大学 | A kind of image characteristic extracting method, device, equipment and readable storage medium storing program for executing |
| CN113033527A (en) * | 2021-05-27 | 2021-06-25 | 北京三快在线科技有限公司 | Scene recognition method and device, storage medium and unmanned equipment |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109711410A (en) | Three-dimensional object rapid segmentation and identification method, device and system | |
| CN110084173A (en) | Number of people detection method and device | |
| Salman et al. | Classification of real and fake human faces using deep learning | |
| CN102722713B (en) | Handwritten numeral recognition method based on lie group structure data and system thereof | |
| CN106529499A (en) | Fourier descriptor and gait energy image fusion feature-based gait identification method | |
| CN111160473A (en) | Feature mining method and device for classified labels | |
| CN111833322B (en) | Garbage multi-target detection method based on improved YOLOv3 | |
| CN105701120A (en) | Method and apparatus for determining semantic matching degree | |
| CN110288605A (en) | Cell image segmentation method and device | |
| CN110059656B (en) | Method and system for classifying white blood cells based on convolution countermeasure generation neural network | |
| CN112215696A (en) | Personal credit evaluation and interpretation method, device, equipment and storage medium based on time sequence attribution analysis | |
| CN114341880A (en) | Techniques for visualizing operation of neural networks | |
| CN110008853A (en) | Pedestrian detection network and model training method, detection method, medium, equipment | |
| CN111339935A (en) | Optical remote sensing picture classification method based on interpretable CNN image classification model | |
| CN108805152A (en) | A kind of scene classification method and device | |
| CN107368526A (en) | A kind of data processing method and device | |
| CN108763096A (en) | Software Defects Predict Methods based on depth belief network algorithm support vector machines | |
| CN112132014A (en) | Target re-identification method and system based on non-supervised pyramid similarity learning | |
| CN109993187A (en) | A kind of modeling method, robot and the storage device of object category for identification | |
| CN112308825A (en) | SqueezeNet-based crop leaf disease identification method | |
| US11615321B2 (en) | Techniques for modifying the operation of neural networks | |
| CN108229505A (en) | Image classification method based on FISHER multistage dictionary learnings | |
| CN117235633A (en) | Mechanism classification method, mechanism classification device, computer equipment and storage medium | |
| CN109934352B (en) | Automatic evolution method of intelligent model | |
| CN103473562B (en) | Automatic training and identifying system for specific human body action |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| TA01 | Transfer of patent application right | ||
| TA01 | Transfer of patent application right |
Effective date of registration: 20191227 Address after: 518109 first floor, building 1b, yunantong Industrial Park, langrong Road, Dalang community, Dalang street, Longhua District, Shenzhen City, Guangdong Province Applicant after: Shenzhen Jinghong Technology Co., Ltd Address before: 201203 Shanghai Pudong New Area Shanghai free trade trial area, 1 spring 3, 400 Fang Chun road. Applicant before: Shanghai Jinghong Electronic Technology Co., Ltd. |
|
| RJ01 | Rejection of invention patent application after publication | ||
| RJ01 | Rejection of invention patent application after publication |
Application publication date: 20181113 |